Terminator for Spam - A Fuzzy Approach Revealed

Transcription

1 Terminator for Spam - A Fuzzy Approach Revealed P.SUDHAKAR 1, G.POONKUZHALI 2, K.THIAGARAJAN 3, K.SARUKESI 4 1 Vernalis systems Pvt Ltd, Chennai Department of Computer Science and Engineering, Rajalakshmi Engineering College, Affiliated to Anna University- Chennai, Tamil Nadu 3 Department of Science and Humanities, KCG College of Technology Affiliated to Anna University-Chennai, Tamil Nadu 4 Hindustan Institute of Technology and Science-Chennai,Tamil Nadu INDIA 1 sudhakar.asp@gmail.com, 2 poonkuzhali.s@rajalakshmi.edu.in, 3 vidhyamannan@yahoo.com, 4 profsaru@gmail.com Abstract - In this information technology world, the highest degree of communication happens through s. Realistically most of the inboxes are flooded with spam s as most of transactions through this internet is affected by Passive attacks and Active attacks. Several algorithms exist in the e-world to defend against spam s. But the fulfilment of accuracy in deducting spam is still oscillating between 80-90%. This clearly shows the necessity for improvement in spam control algorithms on various projections. In this proposed work a new solvent was chosen in the fuzzy word to combat against spam e- mails. Various fuzzy rules are created for spam s and every is enforced to pass through fuzzy rule filter for identifying spam. Results of the each fuzzy rule for the input e- mails are derived to classify the to be spam or consent. Key-Words - , spam, Fuzzy, Fuzzy Control, Fuzzy logic, Spam, Spam deduction, User Attitude. E I. INTRODUCTION -mail spam, known as unsolicited bulk (UBE), junk mail, or unsolicited commercial (UCE), is the practice of sending unwanted messages, frequently with commercial content, in large quantities to an indiscriminate set of recipients. Spam in started to become a problem when the Internet was opened up to the general public in the mid-1990s. It grew exponentially over the following years, and today composes some 80 to 85% of all the in the world, by a "conservative estimate". Pressure to make spam illegal has been successful in some jurisdictions, but less so in others [1]. Spammers take advantage of this fact, and frequently outsource parts of their operations to countries where spamming will not get them into legal trouble. Though, is undoubtedly a very effective method of communication these days but at times it can be quite vexing when one is confronted with so many unwanted s where the recipients miss their important s just because their mailbox space is often eaten up by these unwanted s. The legal status of spam varies from one jurisdiction to another. Spammers collect addresses from chat rooms, websites, customer lists, newsgroups, and viruses which harvest users' address books, and are sold to other spammers. They also use a practice known as " appending" or "epending" in which they use known information about their target (such as a postal address) to search for the target's address. Much of spam is sent to invalid addresses. Spam averages 78% of all sent. According to the Message Anti-Abuse Working Group, the amount of spam was between 88-92% of messages sent in the first half of Most of the inbox is flooded with these Spams which occupies lot of memory space. There are several algorithms available for detecting and filtering spam e- mails. Among the existing algorithms, Bayesian filtering produces best result, still it does not detect all the spam e- mails. Most of the existing algorithms considers content alone for filtering the spam s. To detect all the spam s, existing spam filtering methods has to be enhanced. In this proposed work, a new algorithm is devised with various fuzzy rules and fuzzy variables. Each fuzzy rule will produce Attack Factor values which are consider for arriving result. Each rule Attack Factor value was arrived by comparing input parameter against Black list and White List. Black list contains predetermined spam content. White list contains acceptable contents. This final result from above calculated Attack Factor will decide the input content to be spam or ham or to be sent hold state. The final result of the algorithm was obtained by summing up each rule result value and decision was taken based on the result of the individual rules. 332

2 i. RELATED WORKS Xavier Carreras et al.[2] proposed a Boosting algorithm for Anti Spam filtering. Even though Boosting algorithm delivers good result, possibility of misclassification costs persist inside the AdaBoost learning algorithm. William W. Cohen et al.[3] suggested Speech act theory for filtering. The outcome of Speech act theory highly depend on the learning and this approach shows new projection for classifying spam content. Harris Drucker et al.[7] developed support vector Machines for Spam Categorization. Even though support vector approach outperforms well, switching from training model need user intervention. Addition to that, reply s are considered as no spam. Joes M.Gomez Hidalgo et al.[8] presents a new dimension for spam classification. Nikolos et al.[13] implemented new technique for spam categorization couple with header information and content information. However this system is under research in peer to peer networks. Even though the conceptualization is good, but the practical bottle neck will comes for identification of spam words from the global set. This will take large amount of time as it works with centralized architecture. Peng et al.[9] Proprosed a new system for applying spam filter in distributed environment. The proposed techniques out performs well during implementation of spam filter in the distributed system. But Author fails to state the technique that can be used to identify spam based on content. The technique handled in this approach ( copy rank ) performs based on header rather than header and body content. Wanli et al.[10] projected a new techniques for identifying spam of content type image. But the experimental results shows less confidence on their approach due to misclassification. From the misclassification list, image based classification got highest rank over other text, HTML and non English text classifications. Sadegh et al.[12] follow through a new approach called Bayesian spanning tree with Likelihood function to identify the in the space. From the likelihood classification, Bayesian Spanning Tree outperforms well compared to Navie Bayesian approach by considering precision and F-measure as measurement. Nevertheless Bayesian approach produces high result, still there is a large space to reach 100% accuracy. Bayesian precision measure declares at the maximum of 85% efficiency can be obtained by using Bayesian spanning tree. ii. OUTLINE OF THE DOCUMENT Section 2 composes various fuzzy rules formation for the input parameters to identify the as a spam or consent. Section 3 Implements the fuzzy rules formed over input (s). Section 4 predicts the input and categorize into appropriate buckets. Section 5 proposes results and discussion on the results with future work. II. FUZZY SYSTEM AND FUZZY RULES GENERATION Fuzzy Logic (FL) is a problem-solving control system methodology that lends itself to implementation in systems ranging from simple, small, embedded microcontrollers to large, networked, multi-channel PC or workstation-based data acquisition and control systems. It can be implemented in hardware, software, or a combination of both. FL provides a simple way to arrive at a definite conclusion based upon vague, ambiguous, imprecise, noisy, or missing input information. FL's approach to control problems mimics how a person would make decisions, only much faster. Fuzzy rules have been advocated as a key tool for expressing pieces of knowledge in fuzzy logic i. FUZZYFICATION Input variable : {Sender saddress, Sender_IP, Subject_Words, ContentWords, Attachment} Fuzzy set : {positive, Zero, Negative} Linguistic set : (highpositive, highnegative, Zero} Rule 1: a: IF SenderAddress spammer list AttackFactor=-0.25; b: IF SenderAddress to Ham list AttackFactor=0.25; c : IF Sender Address Spammerlist & Sender address Ham addresslist AttackFactor=0; Explanation: Rule 1.a : If there exist a sender address belongs to spammer list, then Attack Factor of this rule should be set to -0.25; Rule 1.b : If there exist a sender address belongs to Ham list then, Attack Factor of this rule should be set to 0.25; Rule 1.c : If there exist a sender address that doesn t belongs to spammer list and Ham list then, Attack Factor of this rule should be set to 0; 333

3 Rule 2 : a: IF Sender_IP SpammerIPlist AttackFactor= -0.25; b: IF Sender_IP HamIPlist AttackFactor=0.25; c: IF Sender_IP SpammerIPlist & HamIPlist AttackFactot=0; Explanation: Rule 5.a : If all attachment doesn t belong to virus list then, Attack Factor of this rule is set to 1.0; Rule 5.b : If there exist an attachment belongs to virus list, then Attack Factor of this rule is set to -1.0; Explanation: Rule 2.a : If there exists a sender IP address belongs to Spammer list, then Attack Factor of this rule should be set to -0.25; Rule 2.b : if there exists a sender IP address belongs to Ham list, then Attack Factor of this rule was set to 0.25; Rule 2.c : If there exists a sender IP address doesn t belongs to Spammer list and Ham List then Attack Factor of this rule was set to 0; III. FUZZY RULE IMPLEMENTATION Rule 3: a: IF Subject words Spam words AttackFactor= -0.50; b: IF Subjectword Spamwords <AttackFactor< 0.50 Explanation: Rule 3.a: If all Subject words belongs to Spam words then, Attack Factor of this rule should be set to -0.50; Rule 3.b : If there exists a subject word that belongs to spam word then Attack Factor of this rule is varies from to +0.50; Rule 4: a: IF Content words Spamwordlist AttackFactor= -0.50; b: IF Content words Spamwordslist <AttackFactor< 0.50; Explanation: Rule 4.a : If all content words belongs to Spam words then, Attack Factor of this rule should be set to -0.50; Rule 4.b : If there exists an content word that belongs to spam word then Attack Factor of Rule 5 : this rule is varies from to +0.50; a: IF Attachment VirusList AttackFactor=1.0; b: IF Attachment Visuslist AttackFactor= -1.0; Figure 1. Architecture of proposed system When an is arrived, identified fuzzy input parameters are extracted and it is passed to fuzzy system for identification as per Figure1. After Fuzzyfication and Defuzzyfication categorized s are send back to user. Detailed internal follow was shown in Figure 2. Rule 1 was applied on Fuzzy input parameter- Sender address. Based on Rule 1, Sender address was extracted from and compared against the Black list which has spammer address list. If any match was found then, Attack Factor for this rule was set to If sender address was not found in the black list, then it was compared against the White list which contains all good and acceptable addresses. If match was found, then attack factor for this rule was set to If sender address was not found in both Black and White list, then attack factor for this rule was set to 0. Set this rule result in R1. Rule 2 was applied on Fuzzy Input parameter- Sender IP. IP Address of the sender was compared against the IP Address Black List. If match was found, then Rule 2 Attack Factor was set to If not found, then Sender IP Address was compared against White List IP Address. 334

4 If match found then attack factor of Rule 2 was set to If not found then Attack Factor of the Rule 2 was set to 0. Assign resultant value in R2. Rule 3 was applied on Fuzzy input parameter- Subject words. An may contain one or more words in subject line. All subject word and Content words are preprocessed. The pre-process contains the following steps i.e. stemming, stop words elimination and tokenization. Stemming is the process of comparing the root forms of the searched terms to the documents in its database. Stop words elimination is the process of not considering certain words which will not affect the final result. Tokenization is defined as splitting of the words into small meaning full constituents Algorithm for Subject Attack Factor Calculation: Step 1 :Split the Subject content into words say W i where n i 1 Step 2 : assign to T w = n Step 3 :Calculate word Impact Factor W f where W f = 0.5 /T w Step 4 :Perform comparison for each word W i in Black list Step 5 :If match found then update the update W fi = - W f else W fi = W f ; where i <= T w ; Step 6 : Calculate Attach Factor = Step 7 : Calculate R3 = ; From the subject line after pre-processing total words are counted and each word impact on for this rule is calculate. i,e average impact. Now each word are compared against black and white list already available. If it is found in white list then the Attack factor for this word is set as positive. If it is found in black list then the Attack factor was set as negative. Example : Total words = 5 W f = 0.5 / 5 = 0.1 If the word W i is present in While list then the AttackFactor = If the word Wi is present in the Black list then the Attack Factor = Rule 4 was applied on Fuzzy Input variable- ContentWords after Pre-Processing. Every body may contain one or more words. Every words are taken and compared against the Block list words. Following are the Algorithm to compute Attack Factor of Rule 4. Algorithm for Content Attack Factor Calculation: Step 1 :Split the bodycontent to words say W i where i 1 Step 2: Count the total number of words in Bodyand assign to Tw Step 3 : If T w > 0 then continue Step 4. Step 4 :Calculate word impact factor W f where W f = 0.5 /T w Figure 2. Detailed system flow After pre processing all words are taken and compared against the Black list words. Every words impact (Attack Factor) on this subject line was calculated. Following are Algorithm to compute Attack Factor of Rule 3 Step 5 :Perform comparison for each word Wi in Black list Step 6 :If match found then update the update W fi = - W f else W fi = W f ; where i Tw; Step 7 :Calculate Attach Factor = Step 8 : Calculate R4 = ; Rule 5 was applied to calculate Attack Factor for containing attachment. If does not contain Attachment, then Attack Factor was set to zero. If any one of the attachment content was identified in virus list then Attack Factor was set to -1. If none of the content 335

5 was identified in virus list, then Attack Factor was set to 1. Rule 5 result was assigned to R5. Defuzzification: Result value of each was arrived by sum up previous rule results and these results are termed as decision making factors. R1 = R1; R2 = R2 + R1; R3 = R3 + R2; R4 = R4 + R3; R5 = R5 + R4; IV. RESULTS BASED ON USER ATTITUDE AND DISCUSSION Every rule results are obtained and user attitude was taken consideration for categorizing input s. User Attitude was initially configured to take decision based on fuzzy Linguistic set {High Positive, zero, high negative}. High positive users are type of user who strictly restricts spam s. Zero level users are neutral user who does not have restriction. High negative users are more interested in receiving spam s. Following are the possible values of the Linguistic Set High Positive 0.25; Zero = 0; High Negative Decision making for High Negative level user: If user s attitude level was set High Negative and all rule result value is then the e- mail is declared as consent. If user s attitude level was set as High Negative and any one the rule result value is < then the is set to Hold in which user can take final decision. All fuzzy rules are applied over 243 different kind of e- mails using fuzzy input variables: Sender s Address, Sender, Subject Words, Content Words and Attachment. Results of some sample s are distributed in the following tables. Table 1. Results based on Fuzzy Rules with High positive user Attitude E- mail Fuzzy Results Sour ce Rule1 Rule2 Rule3 Rule4 Rule5 Result E Consent E Spam E Hold E Consent E Spam E Hold Following are the decision making process. Decision making for High positive level users: If user s attitude was set as high positive and all applied rules values are > 0.25 then the is declared as consent. If user s attitude was set as high positive and any one of the rule result value various between 0.25 to 0 then the is declared as hold. If user s attitude was set as high positive and any one of the rule value is < 0 then the is set to Spam. Decision making for Zero level user: If user s attitude level was set as Zero and all rule result value is 0 then the is declared as consent. If user s attitude level was set as Zero and any one of rule value is < 0 then the is set to Spam. Figure 3. Graphical representation of Table 1 336

6 Table 2. Results based on Fuzzy user with Zero user Attitude Source Fuzzy Results Rule1 Rule2 Rule3 Rule4 Rule5 Result E Consent E Spam E Consent E Consent E Spam E Consent Figure 5. Graphical representation of Table 3 Table 4 : Results based on Fuzzy Rules with different user Attitudes ES R1 R2 R2 R4 R5 HP Z HN E Consent Consent Consent E Hold Consent Consent E Hold Consent Consent E Spam Spam Spma E Spam Spam Consent E Hold Consent Consent Figure 4. Graphical representation of Table2 From Fig 4 represents nature. If we see any that has negative region then the is set to spam. From the graph we can identify E2 and E5 are spam as it grows in negative region. Table 3. Results based on Fuzzy Rules with high negative user Attitude E- mail Fuzzy Results Sourc e Rule1 Rule2 Rule3 Rule4 Rule5 Result E Spam E Consent E Consent E Spam E Consent E Consent From table 3 the results can be easily predicted that the relaxation of user who intentionally wish to accept spam s, user level was set to So the range from and above the s are categorized as Consent. Below the level, s are categorized as Hold. The same was represented in a graphical manner in Fig 4. ES Source RX Rule X where X varies from 1 to 5 HP High Positive user Attitude Z - Zero user Attitude HN High Negative user Attitude Table 4 Consolidates different user projections on the same with samples. All possible combination results are provided in Appendix different sets of s are taken for evaluation and results are represented in following figures Figure 6. High positive user s attitude Out of 243 s based on high positive user s attitude, 22 e-malis are categorized as Consent, 41 e-malis are 337

7 categorized in Hold state and 180 s are stamped as spam. Acknowledgment The authors would like to thank Dr. Ponnammal Natarajan worked as Former Director Research, Anna University- Chennai,India and currently an Advisor, (Research and Development), Rajalakshmi Engineering College and Dr. K..Ravi, Associate Professor, Department of Mathematics, Sacred Heart College- Tirupattur, India for their intuitive ideas and fruitful discussions with respect to the paper s contribution. REFERENCES Figure 7. Zero level user s attitude Out of 243 s based on zero level user s attitude, 68 e-malis are categorized as Consent and 180 s are stamped as spam. Figure 8. High Negative user s attitude Out of 243 s based on High Negative user s attitude, 112 e-malis are categorized as Consent and 131 s are categorized as Hold. CONCLUSION AND FUTURE WORK In this proposed work, Fuzzy rules are constructed for 5 input parameters namely Sender s Address, Sender_IP, Subject_Words, Content Words and Attachment for common user to deduct the spam s based on the attitude of the user. The proposed simplistic approach out performs in terms of accuracy in deducting spam s than the existing approaches provided the Black list and White lists to be up to date. The proposed approach works only for s having subject and body content as plain text. Future work aims at deducting spam s having images and HTML also. [1] Metrics report [2] Carreras, X. and Mdrquez, L., Boosting trees for anti-spare filtering, In Proc. of RANLP, [3] Cohen, W.W., Learning Rules that Classify ., Proceedings. of the AAAI Spring Symposium on Machine Learning in Information Access, Stanford, California,1996. [4] Cournane, A. and Hunt, R., An Analysis of the Tools Used For the Generation and Prevention of Spam, Computer and Security, Vol. 23, pp , [5] Cox, E., The Fuzzy System Handbook, Academic Press, Second Edition, [6] Daelemans, W., Z. Jakub, K. van der [7] Sloot and A. van den Bosch, TiMBL: Tilburg Memory Based Learner, version 2.0, Reference Guide. ILK,Computational Linguistics, Tilburg University [8] Drucker, H., Wu, D., & Vapnik, V., Support vector machines for Spam categorization. IEEE-NN, Vol. 10, No.5, pp ,1999. [9] Graham, P., Better Baysian Filtering. In Proceedings of Spam Conference [10] Peng Liu, Guangliang Chen, Liang Ye, Weiming Zhong, Proceedings of the 5th WSEAS Int. Conf. On Simulation, Modeling and Optimization, Corfu, Greece, August 17-19, 2005 (pp61-66). [11] Wanli Ma, Dat Tran, Dharmendra Sharma, Sen Li, Proceedings of the 2007 WSEAS International Conference on Computer Engineering and Applications, Gold Coast, Australia, January 17-19, [12] Hidalgo, J. G., Spez, M, and Sanz, E, Combining text and heuristicz for cost-sensitive spam filtering. In Proc. of CONL, [13] Sadegh Kharazmi, Ali FarahmandNejad, Proceeding of the 9th WSEAS Int. Conference on Data Networks, Communications, Computers, Trinidad and Tobago, November 5-7, [14] Nikolaos Korfiatis, Marios Poulosy, Sozon Papavlassopoulos, Proceeding of the WSEAS International Conference on Applied Mathematics, Greece, Aug 19, 2004 ( ). [15] Lee, J., Spam: An escalating attack of the clones, The New York Times, [16] Mayer, C., and Eunjung-Cha, A., Making spam go splat: Sick of unsolicited , [17] businesses are fighting back, The Washington Post, [18] Norvig P. and Russell S., Artificial Intelligence A Modern Approach, Prentice Hall, New Jersey, [19] Nozaki, K., Ishibushi, H. and Tanaka, H., Trainable Fuzzy classification systems based on Fuzzy If-Then-Rules, Proc. IEEE, vol. 1, pp , [20] RFC 822: Standard for the Format of Arpa Internet Text Messages, [21] Sahami, M., Dumais, S., Heckerman, D. and Horvitz, E., A Bayesian Approach to Filtering Junk . In Learning for Text Categorization, AAA1 Workshop, pp , Madison Wisconsin, [22] SpamAssassin,

8 [23] Sudhakar.P, Poonkuzhali.S, Thiagarajan.K and Sarukesi.K., Fuzzy Logic for Spam deduction, Proceedings of the WSEAS 10th International Conference on Applied Computer and Applied Computational Science, Venice, Italy, March 8-10, 2011 ISBN: [24] Poonkuzhali.S, Thiagarajan.K, P.Sudhakar Kishore Kumar.R and Sarukesi.K., Spam Filtering using Signed and Trust Reputation Management, Proceedings of the WSEAS 10th International Conference on Applied Computer and Applied Computational Science, Venice, Italy, March 8-10, 2011 ISBN: P.Sudhakar received Bachelor of Engineering degree in Computer science from Anna University Chennai-India in 2006 and Master of Engineering degree in Computer Science from Anna University Chennai-India in He started his carrier as a Junior software programmer in Vernalis systems Pvt Ltd, Chennai India at 2008 and elevated to Associate software. He also presented various papers in National level conferences and published his research work in International Conferences and Journals. G.Poonkuzhali received B.E degree in Computer Science and Engineering from University of Madras, Chennai, India, in 1998, and the M.E degree in Computer Science and Engineering from Sathyabama University, Chennai, India, in Currently she is pursuing Ph.D programme in the Department of Information and Communication Engineering at Anna University Chennai, India. She has presented and published 10 research papers in international conferences & journals and authored 5 books. She is a life member of ISTE (Indian Society for Technical Education),IAENG (International Association of Engineers), and CSI (Computer Society of India). K.Thiagarajan working as Senior Lecturer in the Department of Mathematics in KCG College of Technology - Chennai-India. He has totally 14 years of experience in teaching. He has attended and presented research articles in 33 National and International Conferences and published one national journal and 26 international journals. Currently he is working on web mining through automata and set theory. His area of specialization is coloring of graphs and DNA Computing. Dr. K. Sarukesi has a very distinguished career spanning of nearly 40 years. He has a vast teaching experience in various universities in India and abroad. He was awarded a commonwealth scholarship by the association of common wealth universities, London for doing Ph.D in UK. He completed his Ph.D from the University of Warwick U.K in the year His area of specializations is Technological Information System. He worked as expert in various foreign universities. He has executed number of consultancy projects. he has been honored and awarded commendations for his work in the field of information technology by the government of TamilNadu. He has published over 40 research papers in international conferences/journals and 40 National Conferences/journals. 339

9 Appendix -1 Source Fuzzy Rules High Positive Zero High Negative Rule1 Rule2 Rule3 Rule4 Rule5 Consent Hol d Spa m Consent Spam Consent E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES Hol d 340

10 E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES 341

11 E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES 342

12 E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES 343

13 E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES 344

14 E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES E YES YES YES 345